def make_birel_matrix(self, relation='children'):
birel = np.zeros((len(
self.graph_structs), self._max_nodes, self._max_bi_relations, 2),
dtype='int32')
for i, gs in enumerate(self.graph_structs):
for j, nid in enumerate(gs.graph.nodes):
nid_token = get_node_token(gs.graph, nid)
for k, rel_nid in enumerate(getattr(gs, relation)[nid]):
rel_token = get_node_token(gs.graph, rel_nid)
try:
# birel[i, j, k, :] = [nid_token, rel_token]
birel[i, j, k, :] = [nid, rel_nid]
except IndexError:
continue
return birel
def make_vocabulary(self):
if self.word2ind is not None and len(self.word2ind) > 1:
logging.info('word2ind already exists ({0} entries). Reusing it. Some entries are: {1}'.format(
len(self.word2ind),
list(self.word2ind.items())[:10]))
return self.word2ind
logging.info('word2ind does not exist. Creating it.')
counter = Counter()
constants = []
special = []
for gs in self.graph_structs:
graph = gs.graph
for nid in graph.nodes:
token = get_node_token(graph, nid)
if get_label(graph, nid, 'type') == 'constant':
constants.append(token)
counter[token] += 1
else:
special.append(token)
counter[token] += 1
logging.info('Most common 10 tokens: {0}'.format(counter.most_common()[:10]))
special = sorted(set(special))
logging.info('Got {0} special tokens: {1}'.format(len(special), special))
constants = sorted(set(constants))
logging.info('Got {0} constant tokens. Some of them are: {1}'.format(
len(constants), constants[:10]))
vocab = special + constants
assert '<unk>' not in vocab
[self.word2ind[w] for w in vocab]
logging.info('word2ind created. Some entries are: {0}'.format(
list(self.word2ind.items())[:10]))
return self.word2ind
def make_treelet_matrix(self, relation='treelet_predicate'):
treelets = np.zeros(
(len(self.graph_structs), self._max_nodes, self._max_treelets, 3),
dtype='int32')
for i, gs in enumerate(self.graph_structs):
for j, nid in enumerate(gs.graph.nodes):
nid_token = get_node_token(gs.graph, nid)
for k, (rel1_nid,
rel2_nid) in enumerate(getattr(gs, relation)[nid]):
rel1_token = get_node_token(gs.graph, rel1_nid)
rel2_token = get_node_token(gs.graph, rel2_nid)
treelets[i, j, k, :] = [
self.word2ind[nid_token], self.word2ind[rel1_token],
self.word2ind[rel2_token]
]
return treelets
def make_vocabulary(self):
counter = Counter()
constants = []
special = []
for gs in self.graph_structs:
graph = gs.graph
for nid in graph.nodes:
token = get_node_token(graph, nid)
if get_label(graph, nid, 'type') == 'constant':
constants.append(token)
counter[token] += 1
else:
special.append(token)
counter[token] += 1
logging.info('Most common 10 tokens: {0}'.format(
counter.most_common()[:10]))
special = sorted(set(special))
logging.info('Got {0} special tokens: {1}'.format(
len(special), special))
constants = sorted(set(constants))
logging.info('Got {0} constant tokens. Some of them are: {1}'.format(
len(constants), constants[:10]))
vocab = special + constants
assert '<unk>' not in vocab
[self.word2ind[w] for w in vocab]
self.word2emb = np.random.uniform(size=(len(self.word2ind), 2))
return self.word2ind
def make_node_inds(self):
node_inds = np.zeros((len(self.graph_structs), self._max_nodes),
dtype='float32')
for i, gs in enumerate(self.graph_structs):
for j, nid in enumerate(gs.graph.nodes):
node_token = get_node_token(gs.graph, nid)
node_inds[i, nid] = self.word2ind[node_token]
return node_inds